Pulse is a high-accuracy, low-latency speech-to-text model built for real-time transcription across 38 languages, with streaming and non-streaming support.
TTFT at 1 concurrency
TTFT at 100 concurrency
Streaming + Non-streaming
Streaming + Non-streaming
Ultra-low latency architecture delivering 64ms TTFT at 1 concurrency and 300ms at 100 concurrent requests — designed for live transcription and conversational AI.
38 languages supported across streaming and non-streaming modes, with automatic language detection and code-switching within a single session.
Built-in redaction of personal and payment card data across both streaming and non-streaming use cases.
Automatic multi-speaker identification across both streaming and non-streaming modes, with per-word and per-utterance speaker labels.
Background noise handling built into the model.
Supports multi-language audio within a single session. Best used by setting the known primary language (e.g. es for Spanish handles English+Spanish automatically).
Pulse STT is evaluated against three open-source datasets — FLEURS, ESB, and WildASR — and one internal English perturbation suite. Word Error Rate (WER) by language. Lower is better. NA = not available or not supported by that provider.
For the full benchmark comparison across every dataset, see the Performance page.
WER on the English subset of FLEURS across providers in streaming mode. Lower is better.
Audio amplitude normalization materially changes WER on FLEURS. Most competitors benchmark on raw FLEURS — which has variable, often low amplitude — without normalizing peak audio to −10 dBFS. This makes some models look much better than they actually are. Pulse is stable across all amplitude regimes.
WER across seven Hindi datasets covering read speech, conversational speech, telephony / contact-center audio, and noise-augmented variants. Compared against IndicWhisper, Sarvam Saaras v3, and Deepgram Nova-3. Lower is better.
For the full breakdown including training-data and evaluation-protocol notes, see the Performance page.
A Hugging Face benchmark suite aggregating 9 English speech datasets across diverse domains (audiobooks, parliament, meetings, finance, etc.) to test STT generalization. Lower WER is better.
Evaluated on the open-source Hugging Face ESB datasets. Numbers from internal evaluation.
An open-source robustness benchmark designed to stress-test STT under real-world degraded conditions: clipping, far-field capture, background noise, phone codec compression, reverberation, and accented speech. Lower WER is better. n/a = not supported by that provider.
Evaluated on the open-source WildASR dataset. Numbers from internal evaluation.
Not a public dataset. The English audio is sliced by perturbation type (Noise, Silence, Telephony 911, Boundary, Disfluency, Long Audios, Repetition, Entity, Accent, Emotion, Speaker Diversity, Speed, Pitch, Volume, Audio Quality) to isolate model weaknesses. Lower WER is better.
Not a public dataset. Hindi audio sliced by perturbation type to isolate model weaknesses. Lower WER is better except for Entity EDR where higher is better (↑).
When the language of the audio is known in advance, always set it explicitly rather than relying on automatic detection. This yields better transcription accuracy because the model can optimize directly for that language without needing to first identify it.
For example, setting the language parameter to es (Spanish) tells the model to expect Spanish audio, which also handles English+Spanish code-switching scenarios. This produces more accurate outputs compared to using multi-eu or multi.
When to use multi-eu or multi:
multi-eu for European-language input; use multi only for truly mixed multilingual audioPulse must not be used for:
Additionally: